CN-122024161-A - Electric power inspection self-adaptive method and system based on cloud edge cooperation and layered skill migration

CN122024161ACN 122024161 ACN122024161 ACN 122024161ACN-122024161-A

Abstract

The invention discloses a power inspection self-adaptive method and system based on Yun Bian cooperation and layered skill migration, according to the method, a multi-mode large model core base and a large scene expert LoRA adapter are deployed at the cloud end, a lightweight execution base is deployed at the edge side, and a centralized skill atom library is constructed. And (3) taking the corresponding cloud expert model as a teacher and the cloud copy of the edge base as a student, generating an edge lightweight small LoRA adapter through layered knowledge distillation, and binding and registering after verification. After the inspection task is received, analyzing and generating a skill execution diagram, distributing cloud edge execution positions by combining hard constraint and real-time state, loading an adapter execution reasoning and shunting results by edge equipment in stages. Collecting a multi-source sample, calculating comprehensive value scores, screening a high-value sample, optimizing a cloud expert model, redistilling to generate an edge adapter, and performing gradual deployment after verification to realize continuous evolution of the system. The cloud edge knowledge dynamic programming method solves the problems that cloud edge coordination is insufficient, intelligent capability is difficult to dynamically program, and cloud edge knowledge cannot form closed loop evolution.

Inventors

XU XING
ZHANG ZUWEI
ZENG JIAN
GAO LIPING
ZHAO JIANBIN
ZHANG BO
ZHANG PENGFEI
LIU YADUO
ZHAO XIAOXIANG
Geng Mingxi
YANG CHEN
WANG SHAOYING

Assignees

国网河北省电力有限公司信息通信分公司
北京西清能源科技有限公司

Dates

Publication Date: 20260512
Application Date: 20260202

Claims (10)

1. The electric power inspection self-adaptive method based on cloud edge cooperation and layered skill migration is characterized by comprising the following steps: deploying a multi-mode large model as a core base and deploying a scene expert large LoRA adapter at the cloud, deploying a lightweight model as an execution base at the edge, decoupling and packaging the patrol capacity into standardized atomic skills based on a unified capacity abstract strategy, and constructing a centralized skill atomic library; Based on the definition of the set atomic skills requiring edge deployment in the centralized skill atomic library, taking a cloud scene expert model to which the set atomic skills belong as a teacher, taking cloud copies of an edge lightweight base as students, and training through a hierarchical knowledge distillation technology to generate an edge lightweight small LoRA adapter; According to hard constraint in skill metadata, combining a real-time network state, edge equipment load and skill performance parameters, distributing cloud edge execution positions for flexible scheduling skill nodes in the directed acyclic skill execution graph to obtain an optimized execution graph, dynamically loading the edge lightweight small LoRA adapter by edge equipment according to the optimized execution graph in stages, and executing pipeline reasoning and shunting results; Collecting edge low confidence samples, cloud edge discrimination difference samples and new defect mode samples generated in a pipeline reasoning process, calculating to obtain comprehensive value scores of the samples, screening high-value samples to form an evolution training set based on the comprehensive value scores, optimizing cloud expert high LoRA adapters through the evolution training set to obtain optimized expert models, carrying out knowledge distillation again based on the optimized expert models to generate new-generation edge lightweight small LoRA adapters, and gradually deploying the new-generation edge lightweight small LoRA adapters to edge equipment after verification to realize continuous evolution of a system.
2. The electric power inspection self-adaption method based on cloud edge coordination and layered skill migration according to claim 1, wherein the multi-modal large model is characterized in that the expression of a training loss function is as follows by performing joint optimization training on a large-scale image-text pair data set: ; In the formula, Pre-training loss for multi-modal large models; to a large-scale universal multi-modal data set Taking expectations of samples (x, y), wherein x is input data, and y is corresponding text description; Modeling the penalty for the mask language; loss for image-text contrast; Loss for image-text matching; The light model is derived from a cloud large model through a knowledge distillation technology, and the basic distillation loss function expression of the light model is as follows: ; In the formula, The distillation loss is based on a cloud edge model; to a non-label data set Taking a desire for an input sample x of (a); Outputting middle characteristics of input x for the cloud large model; And outputting the intermediate characteristics of the input x for the edge lightweight model.
3. The power inspection adaptive method based on cloud edge collaboration and layered skill migration of claim 2, wherein in the process of generating the edge lightweight small LoRA adapter through training of the layered knowledge distillation technology, the layered knowledge distillation technology trains to generate the edge lightweight small LoRA adapter by adopting a multi-objective combined loss function, wherein the multi-objective combined loss function comprises a KL divergence loss, an intermediate feature alignment loss and a task specific supervision loss; the expression of the multi-objective combined loss function is as follows: ; In the formula, Loss for multi-objective combinations; Loss of KL divergence; a dimension projection function; Loss of supervision specific to a task; alignment loss for intermediate features; 、、 Are all weight coefficients; The probability distribution is output for the student model; the probability distribution is output for the teacher model; outputting characteristics of an intermediate layer of the student model; outputting characteristics of the middle layer corresponding to the teacher model; A prediction label output by the teacher model; predictive labels output for student models.
4. The adaptive power inspection method based on cloud edge collaboration and layered skill migration as claimed in claim 3, wherein in the process of distributing cloud edge execution positions for flexible scheduling skill nodes in the directed acyclic skill execution graph, adaptive optimization is performed on weight parameters of a scheduling strategy by adopting an online learning strategy based on a sliding window, and utility contribution adjustment amounts of weight components are calculated in the sliding window: ; wherein W is the size of the sliding window; the normalized value of the jth utility component in the ith scheduling is obtained; Is the average value in the window; is self-adaptive learning rate; The method comprises the steps of constructing a scheduling strategy-environment state mapping library, associating a system state mode with an empirically verified optimization weight, directly loading the optimization weight corresponding to a set mode to perform quick response when detecting that the matching degree of a current state and the set mode in the scheduling strategy-environment state mapping library exceeds a set matching degree threshold value, and triggering a weight resetting mechanism when the comprehensive dispatching utility of continuous set times is lower than the set utility threshold value or the system state is suddenly changed: ; In the formula, Is the system weight; Default weights are preset; is a weight snapshot of the most recent steady state.
5. The adaptive power inspection method based on cloud edge collaboration and layered skill migration according to claim 4, wherein the calculation formula of the comprehensive value score is as follows: ; In the formula, Is a composite value score; Is a sample; for measuring uncertainty of the model to its predictions; The cloud edge identification method is used for identifying cloud edge identification conflicts; for evaluating the novelty of the sample in the feature space; 、、 Are all weight coefficients; verification of the new generation edge lightweight small LoRA adapter includes performance verification, compatibility verification, and robustness verification.
6. The adaptive power inspection method based on cloud edge collaboration and layered skill migration according to claim 5, wherein in the process of deploying the new generation edge lightweight small LoRA adapter to an edge device through the progression, a deployment progress control function expression is: ; In the formula, Is the total number of devices; Controlling the deployment rate; Is the start time; After deployment, carrying out quantitative evaluation on the overall effect of the skill evolution by calculating an evolution gain index, wherein the calculation formula of the evolution gain index is as follows: ; In the formula, The amplitude is improved for the accuracy; The rate of change of resource overhead; for improved performance.
7. The electric power inspection self-adaptive system based on cloud edge cooperation and layered skill migration adopts the electric power inspection self-adaptive method based on cloud edge cooperation and layered skill migration as claimed in any one of claims 1-6, and is characterized by comprising the following steps: The centralized skill atomic library construction module is used for deploying a multi-mode large model as a core base and deploying a scene expert large LoRA adapter at the cloud end, deploying a lightweight model as an execution base at the edge, and decoupling and packaging the patrol ability into standardized atomic skills based on a unified ability abstraction strategy to construct a centralized skill atomic library; The edge light small LoRA adapter generation module is used for generating an edge light small LoRA adapter by training through a hierarchical knowledge distillation technology based on the definition of the set atomic skills requiring edge deployment in the centralized skill atomic library, taking a cloud scene expert model to which the set atomic skills belong as a teacher, taking cloud copies of an edge light base as students, binding the edge light small LoRA adapter with the set atomic skills after performance verification, and registering the edge light small LoRA adapter to the centralized skill atomic library; The system comprises a directed acyclic skill execution graph generation and execution module, a cloud edge execution position distribution module, a processing module and a processing module, wherein the directed acyclic skill execution graph generation and execution module is used for analyzing a patrol task based on the skill atom library after receiving the patrol task and generating a directed acyclic skill execution graph; The closed loop feedback and iteration optimization module is used for collecting edge low confidence samples, cloud edge discrimination difference samples and new defect mode samples generated in the pipeline reasoning process, calculating to obtain comprehensive value scores of the samples, screening high-value samples to form an evolution training set based on the comprehensive value scores, optimizing the cloud end scenerization expert large LoRA adapter through the evolution training set to obtain an optimized expert model, carrying out knowledge distillation processing again based on the optimized expert model to generate a new-generation edge lightweight small LoRA adapter, verifying the new-generation edge lightweight small LoRA adapter, and gradually deploying the new-generation edge lightweight small LoRA adapter to edge equipment to realize continuous evolution of a system.
8. The electric power inspection self-adaptive system based on cloud edge collaboration and layered skill migration according to claim 7, wherein in the centralized skill atom library construction module, the multi-modal large model performs joint optimization training on a large-scale image-text pair data set, and the expression of a training loss function is as follows: ; In the formula, Pre-training loss for multi-modal large models; to a large-scale universal multi-modal data set Taking expectations of samples (x, y), wherein x is input data, and y is corresponding text description; Modeling the penalty for the mask language; loss for image-text contrast; Loss for image-text matching; The light model is derived from a cloud large model through a knowledge distillation technology, and the basic distillation loss function expression of the light model is as follows: ; In the formula, The distillation loss is based on a cloud edge model; to a non-label data set Taking a desire for an input sample x of (a); Outputting middle characteristics of input x for the cloud large model; And outputting the intermediate characteristics of the input x for the edge lightweight model.
9. The power inspection adaptive system based on cloud edge collaboration and layered skill migration of claim 8, wherein in the edge lightweight small LoRA adapter generation module, in generating the edge lightweight small LoRA adapter through training of the layered knowledge distillation technique, the layered knowledge distillation technique trains to generate the edge lightweight small LoRA adapter with a multi-objective combined loss function comprising KL divergence loss, intermediate feature alignment loss, and task specific supervision loss; the expression of the multi-objective combined loss function is as follows: ; In the formula, Loss for multi-objective combinations; Loss of KL divergence; a dimension projection function; Loss of supervision specific to a task; alignment loss for intermediate features; 、、 Are all weight coefficients; The probability distribution is output for the student model; the probability distribution is output for the teacher model; outputting characteristics of an intermediate layer of the student model; outputting characteristics of the middle layer corresponding to the teacher model; A prediction label output by the teacher model; predictive labels output for student models.
10. The power inspection self-adaptive system based on cloud edge collaboration and layered skill migration as claimed in claim 9, wherein in the generation and execution module of the directed acyclic skill execution graph, in the process of distributing cloud edge execution positions for flexible dispatch skill nodes in the directed acyclic skill execution graph, a sliding window-based online learning strategy is adopted to perform self-adaptive optimization on weight parameters of a dispatch strategy, and utility contribution adjustment amounts of weight components are calculated in the sliding window: ; wherein W is the size of the sliding window; the normalized value of the jth utility component in the ith scheduling is obtained; Is the average value in the window; is self-adaptive learning rate; The method comprises the steps of constructing a scheduling strategy-environment state mapping library, associating a system state mode with an empirically verified optimization weight, directly loading the optimization weight corresponding to a set mode to perform quick response when detecting that the matching degree of a current state and the set mode in the scheduling strategy-environment state mapping library exceeds a set matching degree threshold value, and triggering a weight resetting mechanism when the comprehensive dispatching utility of continuous set times is lower than the set utility threshold value or the system state is suddenly changed: ; In the formula, Is the system weight; Default weights are preset; A weight snapshot that is the most recent steady state; in the closed loop feedback and iterative optimization module, the calculation formula of the comprehensive value score is as follows: ; In the formula, Is a composite value score; Is a sample; for measuring uncertainty of the model to its predictions; The cloud edge identification method is used for identifying cloud edge identification conflicts; for evaluating the novelty of the sample in the feature space; 、、 Are all weight coefficients; verification of the new generation edge lightweight small LoRA adapter includes performance verification, compatibility verification, and robustness verification; in the closed loop feedback and iteration optimization module, in the process of deploying the new generation of edge lightweight small LoRA adapter to edge equipment through the progression, a deployment progress control function expression is: ; In the formula, Is the total number of devices; Controlling the deployment rate; Is the start time; After deployment, carrying out quantitative evaluation on the overall effect of the skill evolution by calculating an evolution gain index, wherein the calculation formula of the evolution gain index is as follows: ; In the formula, The amplitude is improved for the accuracy; The rate of change of resource overhead; for improved performance.

Description

Electric power inspection self-adaptive method and system based on cloud edge cooperation and layered skill migration Technical Field The invention relates to the technical field of intelligent inspection and multi-mode artificial intelligence of an electric power system, in particular to an electric power inspection self-adaption method and system based on cloud edge cooperation and layered skill migration. Background In the field of intelligent inspection of power systems, the prior art mainly forms a binary deployment pattern of a cloud heavy model and an edge light model. The cloud side generally adopts a large-scale multi-mode artificial intelligent model, an expert model with complex defect recognition and small target detection capability is constructed by fusing multi-source data such as visual images, infrared thermal images and partial discharge signals, a core support is provided for high-precision inspection analysis, and a lightweight model subjected to cutting, quantization or knowledge distillation treatment is deployed on the edge side, so that calculation force and power consumption constraint of mobile equipment such as an unmanned plane and an inspection robot are adapted, and the on-site real-time response requirement is met. Meanwhile, partial systems try to realize cloud-edge interaction through data uploading, result feedback and other modes, and gradually advance the mode evolution from periodic inspection to state inspection and predictive maintenance. However, the prior art still has significant bottlenecks that on one hand, cloud expert models are large in parameter scale, high in calculation complexity, depend on high-performance hardware clusters, cannot be directly deployed on edge equipment, so that 'capability is in the cloud and calculation power of being in the edge' is mismatched, an advanced algorithm is difficult to land on the first line, and an edge lightweight model is high in omission rate of key defects such as micro cracks and early leakage due to insufficient feature extraction capability and multi-mode fusion loss, weak in generalization capability and easy to cause false alarm due to environmental interference. On the other hand, the cloud edge lacks an efficient knowledge transfer and collaborative learning mechanism, the diagnosis experience accumulated by the cloud edge cannot be effectively sunk to the edge, new scenes and new defects encountered by the edge are difficult to be quickly and reversely fed into the cloud model for optimization, the model updating iteration period is long, the mechanism is stiff, meanwhile, the existing system mostly adopts a fixed processing flow, lacks dynamic resource scheduling capability based on task characteristics and real-time states, has low resource utilization efficiency, is difficult to adapt to the requirements of novel power equipment expansion and inspection mode upgrading, and becomes a key obstacle for restricting intelligent continuous development of inspection. Therefore, there is a need for an adaptive method for power inspection based on cloud edge coordination and layered skill migration to solve the problems of difficult landing of cloud end capability, low edge model accuracy and the like in the prior art. Disclosure of Invention Therefore, the invention provides the self-adaptive method and the system for electric power inspection based on cloud edge coordination and layering skill migration, which solve the problems that the cloud edge coordination is insufficient, the intelligent capability is difficult to dynamically arrange, and the cloud edge knowledge cannot form closed-loop evolution calculation power mismatch in electric power inspection, so that the respective advantages of the cloud edge are brought into full play through a more elastic coordination mechanism, and the preprocessing and pre-screening capability of the edge is enhanced through a flexible arrangement and dynamic updating mechanism, thereby better matching with cloud work. In order to achieve the purpose, the invention provides the following technical scheme that the electric power inspection self-adaptive method based on cloud edge cooperation and layered skill migration is characterized by comprising the following steps: deploying a multi-mode large model as a core base and deploying a scene expert large LoRA adapter at the cloud, deploying a lightweight model as an execution base at the edge, decoupling and packaging the patrol capacity into standardized atomic skills based on a unified capacity abstract strategy, and constructing a centralized skill atomic library; Based on the definition of the set atomic skills requiring edge deployment in the centralized skill atomic library, taking a cloud scene expert model to which the set atomic skills belong as a teacher, taking cloud copies of an edge lightweight base as students, and training through a hierarchical knowledge distillation technology to generate an edge lightweight sma